Novel whole-genome analysis methods for Alzheimer's risk prediction

Despite recent advancements, existing genetic risk prediction models (GRPMs) for late-onset Alzheimer's Disease (LOAD) lack sufficient discrimination ability to support clinical applications. Given the lack of treatments that meaningfully affect disease progression once symptoms have manifested and the socioeconomic consequences at stake, there is a serious unmet need for more accurate GRPMs able to assess a patient's LOAD risk in middle age or earlier, before presymptomatic neurodegeneration begins. Additionally, the lack of diversity in traditional GRPMs further exacerbates inequities in health care for non-Europeans. This project builds off of our previous work in which we produced a GRPM for LOAD using single SNPs and epistatic interactions between SNPs in an ensemble model with polygenic risk scores. In this phase we will develop methods for cross-ancestry SNP analysis to select features to be included in a LOAD GRPM that can be generalized to individuals of any ancestry.
The ADSP data will be incorporated with data from the Alzheimer's Disease Neuroimaging Initiative, the Alzheimer's Disease Sequencing Project, the UK BioBank and the Framingham Heart Study. We will develop methods for association testing for homogeneous cross-ancestry SNP effects, followed by fine-mapping using diverse subjects to identify genetic features with functional relevance. We will also create and implement methods for association testing for heterogeneous cross-ancestry SNP effects for identification of epistatic interactions across ancestries. Finally we will train a cross-ancestry AD risk model using a cross-validation framework and replicate our findings in an independent cohort. The model will include age, sex, APOE genotype, and ancestry PCs as covariates and we will test gradient boosting, deep learning, super learning, and ensemble modeling methods.

There are no treatments that significantly slow progression of Alzheimer’s Disease (AD) once symptoms manifest, making early intervention crucial to reduce the burden of this disease. A genetic risk prediction model (GRPM) for determining AD risk early in life, would allow early intervention, life planning, and improved patient stratification for clinical trials. Despite advancements, GRPMs for AD lack sufficient discrimination to support such applications. To address this need, Parabon developed a GRPM using machine learning that is able to predict an individual's risk of developing AD at any age that achieves state-of-the-art prediction accuracy. However, like most genetic risk scores, it was built using European subjects and thus has reduced accuracy in non-Europeans. The goal of this project is to enhance our modeling pipeline to identify genetic variants and interactions both across and within ancestral groups, then build a predictive model and test it in an independent replication set.